first commit

README.md

@@ -0,0 +1,33 @@
+# Gaussian Constrained Attention Network for Scene Text Recognition
+## Introduction
+Implementation of the paper "Gaussian Constrained Attention Network for Scene Text Recognition" (Under Review)
+## How to use
+### Install
+```
+pip3 install -r requirements.txt
+```
+### Train
+* <b> Data prepare</b>  
+LMDB format is suggested. refer [here](https://github.com/bgshih/crnn/blob/master/tool/create_dataset.py) to generate data in LMDB format.
+
+* <b> Run</b>  
+
+	```
+	python3 train.py --checkpoints /path/to/save/checkpoints --train_data_dir /path/to/your/train/LMDB/data/dir --test_data_dir /path/to/your/validation/LMDB/data/dir -g "0" --train_batch_size 128 --val_batch_size 128 --aug True --att_loss_type "l1" --att_loss_weight 10.0
+	```  
+	More hyper-parameters please refer to [config.py](https://github.com/Pay20Y/GCAN/blob/master/config.py)  
+
+### Test
+
+* Download the pretrained model from [BaiduYun](https://pan.baidu.com/s/1hY374pvtDtgeBUPsG7R5ew) (key:w14k)
+* Download the benchmark datasets from [BaiduYun](https://pan.baidu.com/s/1Z4aI1_B7Qwg9kVECK0ucrQ) (key: nphk) shared by clovaai in this [repo](https://github.com/clovaai/deep-text-recognition-benchmark)
+
+```
+python3 test.py --checkpoints /path/to/the/pretrained/model --test_data_dir /path/to/the/evaluation/benchmark/lmdb/dir -g "0"
+```  
+
+## Experiments on benchmarks
+
+|  IIIT5K | IC13  |  IC15  |  SVT | SVTP  |  CUTE  |
+|:-------:|:-----:|:------:|:----:|:-----:|:------:|
+|  94.4   | 93.3  |  77.1  | 90.1 |  81.2 |  85.6  |

config.py

@@ -0,0 +1,70 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+
+parser = argparse.ArgumentParser(description="Softmax loss classification")
+
+# Data
+parser.add_argument('--train_data_dir', nargs='+', type=str, metavar='PATH',
+                    default=[None])
+parser.add_argument('--train_data_gt', nargs='+', type=str, metavar='PATH',
+                    default=[None])
+parser.add_argument('--test_data_dir', type=str, metavar='PATH',
+                    default=None)
+parser.add_argument('--test_data_gt', type=str, metavar='PATH',
+                    default=None)
+parser.add_argument('-b', '--train_batch_size', type=int, default=128)
+parser.add_argument('-v', '--val_batch_size', type=int, default=128)
+parser.add_argument('-j', '--workers', type=int, default=2)
+parser.add_argument('-g', '--gpus', type=str, default='1')
+parser.add_argument('--height', type=int, default=48, help="input height")
+parser.add_argument('--width', type=int, default=160, help="input width")
+parser.add_argument('--aug', type=bool, default=False, help="using data augmentation or not")
+parser.add_argument('--keep_ratio', action='store_true', default=True,
+                    help='length fixed or lenghth variable.')
+parser.add_argument('--voc_type', type=str, default='ALLCASES_SYMBOLS',
+                    choices=['LOWERCASE', 'ALLCASES', 'ALLCASES_SYMBOLS'])
+parser.add_argument('--num_train', type=int, default=-1)
+parser.add_argument('--num_test', type=int, default=-1)
+
+# Model
+parser.add_argument('--max_len', type=int, default=30)
+parser.add_argument('--encoder_sdim', type=int, default=512,
+                    help="the dim of hidden layer in encoder.")
+parser.add_argument('--encoder_layers', type=int, default=2,
+                    help="the num of layers in encoder lstm.")
+parser.add_argument('--decoder_sdim', type=int, default=512,
+                    help="the dim of hidden layer in decoder.")
+parser.add_argument('--decoder_layers', type=int, default=2,
+                    help="the num of layers in decoder lstm.")
+parser.add_argument('--decoder_edim', type=int, default=512,
+                    help="the dim of embedding layer in decoder.")
+parser.add_argument("--att_loss_type", type=str, default='l1')
+parser.add_argument("--att_loss_weight", type=float, default=10.)
+
+# Optimizer
+parser.add_argument('--lr', type=float, default=0.0008, help="learning rate of new parameters, for pretrained ")
+parser.add_argument('--weight_decay', type=float, default=0.9) # the model maybe under-fitting, 0.0 gives much better results.
+parser.add_argument('--decay_iter', type=int, default=100000)
+parser.add_argument('--decay_bound', nargs='+', type=int, default=[180000, 240000]) # [120000, 150000] (800K)
+# parser.add_argument('--lr_stage', nargs='+', type=float, default=[0.001, 0.1, 0.01])
+parser.add_argument('--lr_stage', nargs='+', type=float, default=[0.0001, 0.00001, 0.000001])
+parser.add_argument('--decay_end', type=float, default=0.00001)
+parser.add_argument('--grad_clip', type=float, default=-1.0)
+parser.add_argument('--iters', type=int, default=3000000)
+parser.add_argument('--decode_type', type=str, default='greed')
+
+parser.add_argument('--resume', type=bool, default=False)
+parser.add_argument('--pretrained', type=str, default='', metavar='PATH')
+parser.add_argument('--checkpoints', type=str, default='./checkpoints/', metavar='PATH')
+parser.add_argument('--log_iter', type=int, default=100)
+parser.add_argument('--summary_iter', type=int, default=1000)
+parser.add_argument('--eval_iter', type=int, default=2000)
+parser.add_argument('--save_iter', type=int, default=2000)
+parser.add_argument('--vis_dir', type=str, metavar='PATH')
+
+def get_args(sys_args):
+  global_args = parser.parse_args(sys_args)
+  return global_args